Professional Documents
Culture Documents
76 (2014) 28105-28108
Control Engineering
Elixir Control Engg. 76 (2014) 28105-28108
Tele:
E-mail addresses: a.moosavian@ut.ac.ir
© 2014 Elixir All rights reserved
28106 A. Moosavian et al./ Elixir Control Engg. 76 (2014) 28105-28108
Journal bearings are applied to carry radial loads, for The sequence of N complex numbers x0,...,xN−1 is
example, to support a crankshaft of engine. Journal bearings are transformed into the sequence of N complex numbers X0,...,XN−1
considered to be sliding bearings as opposed to rolling bearings by the DFT according to the formula:
such as ball bearings. Despite this categorization, a crankshaft N −1 −
2π i
kn , k = 0,..., N − 1 (1)
spinning within a journal bearing is actually separated from the X k = ∑ xn e N
b∈R (2)
where the vector w defines the hyperplane, b is a scalar
threshold. The equations for class 1 and -1, respectively, are
( w.x ) + b = 1 (3)
and
( w.x) + b = −1 (4)
The following decision function holds good for all data points
Figure 2. Diagram of diagnostic procedure belonging to either 1 or -1:
Signal Processing f ( x ) = sign (( w.x) + b) (5)
The great advances in vibration analysis in recent years are
the extension in signal processing techniques, for vibration The optimal hyperplane can be obtained by solving the
diagnostics of journal-bearing defects [6-7-8-9]. The frequency optimization problem:
domain spectrum is more useful in identifying the exact nature minimize: 1 2 (6)
τ ( w) = w
of fault in the journal-bearings. Although there are various 2
techniques, the analysis of vibration signals is often based on the subject to: yi (( w.xi ) + b) ≥ 1 , i = 1,..., l (7)
Fast Fourier Transform (FFT) [10, 23]. where l is the number of training sets. The solution of the
The FFT is a faster version of the Discrete Fourier
Transform (DFT). The Fourier transform converts waveform constrained quadratic programming (QP) optimization problem
data in the time domain into the frequency domain. The Fourier can be obtained as
transform implements this by breaking down the original time- w = ∑ vi xi (8)
based waveform into a series of sinusoidal terms, each with a
unique magnitude, frequency, and phase.
28107 A. Moosavian et al./ Elixir Control Engg. 76 (2014) 28105-28108
where x are SVs obtained from training. Putting (8) in (5), the
i
decision function is obtained as follows:
l (9)
f ( x ) = sign(∑ vi ( x.xi ) + b)
i =1
In cases where the linear hyperplane in input spaces will not be
enough to separate two classes exactly, it is possible to construct
a hyperplane that allows linear separation in the higher Figure 3. Frequency spectrum of normal journal-
dimension (feature space). In SVMs, this is achieved through the bearing
use of a mapped function φ ( x ) that transforms the data from an
N -dimensional input space to Q -dimensional feature space.
Substituting the mapping in (9) gives
l (10)
f ( x) = sign(∑ vi (φ ( x ).φ ( xi )) + b)
i =1
The mapping into higher-dimensional feature space is relatively
severe computation. A kernel can be used to perform this Figure 4. Frequency spectrum of defective journal-bearing
mapping. It means that the transformation can be replaced by an The data mining process is significantly effective in
equivalent kernel function. This helps in reducing the classification results. 30 features were extracted from the
computational load. The kernel function K ( x, y ) is defined as vibration signals in frequency-domain. Then 2 superior features
K ( x, y ) = φ ( x ).φ ( y ) (11) were selected by the Stepwise Backward Selection technique
The decision function is accordingly modified as with p value of 0.05 and linear model. The dataset consisting of
l (12) 2 features and 40 samples were divided into two subsets,
f ( x) = sign(∑ vi K ( x, xi ) + b) training set and test set. The SVM classifier was trained by
i =1 training dataset, and then their performance was exactly
There are different kernel functions like polynomial, quadratic, estimated by the test dataset.
sigmoid, RBF, multilayer perceptron. For SVM, the target values were specified as 1 and -1,
In below, RBF kernel, given by Eq. (13) is described [25]. respectively, representing normal and defective journal-bearing.
− x− y
2 , 0 < σ < +∞ (13) The SVM with RBF kernel width of 1 (σ) and QP method were
K ( x, y ) = exp( 2
) applied for this work. Figure 5 shows the classification result of
2σ journal-bearing conditions. The train and test success of the
For SVM, QP method and the RBF kernel with σ=1 are used. SVM was 95% and 93%. So, it can be found that the
K-Nearest Neighbor performance of proposed procedure was perfect in the fault
Some training samples are used for train the KNN rule. K diagnosis of main engine journal-bearing.
nearest neighbour rule holds position of training samples and
their class. When decision about new incoming data is needed,
distance between query data and training samples is being
calculated. Based on the defined threshold for the rule (it is the
K number), K samples with least distances is selected and the
class with more samples inbound is the result. In the other word,
for example if there is 2 or 3 features for a classification
situation, position of training samples and input sample can be
visualized on 2D and 3D
Cartesian coordinates. Process to find result is like to draw a
circle (Sphere) centred on input location and increase radius
until k samples are embed inside the circle (sphere) and then a
class with more samples inbound is the result. Figure 5. The classification of normal and defective journal-
Without prior knowledge, the KNN classifier usually bearing by using SVM
applies Euclidean distances as the distance metric. However, this KNN is a classifier that its accuracy is always 100% on
simple and easy-to-implement method can still yield competitive training dataset, because the KNN holds the position of training
results even compared to the most sophisticated machine dataset and their class during the classification process. The test
learning methods [26]. The Euclidean distance between point p success of the KNN just depends on the K value. The K value of
and q is the length of the line between them. 1 was used for the KNN. The performance of KNN was 89.84%.
Results and Discussion By attention to the results, it can be found that the
The aim of this paper is the journal-bearing fault diagnosis performance of SVM was significantly better than KNN in the
of engine. To fulfill this purpose, FFT signal processing, data fault detection of main engine journal-bearing.
mining technique, SVM and KNN classifiers were used. Conclusion
Figure 3 and 4 shows the spectrum analysis of the normal A procedure was presented for the fault diagnosis of main
and faulty journal-bearing. It is obvious that the overall vibration journal-bearing in IC engine by using the data mining technique
amplitude of journal-bearing for faulty condition is more. and Support vector Machine classifier. The Fast Fourier
Transform technique is applied for processing signals. The
features of the generator output signal at different journal-
28108 A. Moosavian et al./ Elixir Control Engg. 76 (2014) 28105-28108
bearing conditions are extracted by the statistical and vibration [11] Gunn SR, Brown M, Bossley KM. Network performance
parameters. To further diminish the dimensionality of the assessment for neurofuzzy data modeling. Intelligent Data
extracted feature vectors and enhance the model performance, Analysis 1997;1208 of Lecture Notes in Computer Science: pp
stepwise backward selection was used. Finally, the feature 313-323.
vectors are fed into the SVM and KNN. The performance of [12] Shirvastava SK, Gharde SS. Support Vector Machine for
SVM is calculated by applying RBF kernel and Quadratic Handwritten Devanagari Numeral Recognition. International
Programming method. KNN with K=1 and Euclidean distance Journal of Computer Applications 2010;7(11).
was used. The results of recognition rates are used to estimate [13] Shigeo A. Support Vector Machine for Pattern
the work efficiency in the proposed system. The FFT and SVM Classification. Springer. 2005: pp 237-245.
techniques can provide good diagnosis efficiency for fault [14] Vapnik V. Statistical Learning Theory. Springer, NewYork.
journal-bearing positions identification and journal-bearing fault 1998.
conditions classification, respectively. The total recognition [15] Cortes C, Vapnik V. Support-vector network. Machine
rates are over 93%. Therefore, the results indicated that the Learning 1995;20: pp 273-297.
proposed system can provide an accurate and automatic [16] Bagheri B, Ahmadi H, Labbafi R. Implementing discrete
diagnosis technique for the journal-bearing of IC engines. wavelet transform and artificial neural networks for acoustic
Acknowledgement condition monitoring of gearbox. Elixir Mech. Engg 2011;35: pp
The authors would like to thank Tehran University for its 2909-2911.
support to this research. [17] Niu G, Widodo A, Son JD, Yang BS. Decision-level fusion
References based on wavelet decomposition for induction motor fault
[1] Widodo A, Yang BS. Support vector machine in machine diagnosis using transient current signal, Expert Systems with
condition monitoring and fault diagnosis. Mechanical Systems Applications 2008:35: pp 918-928.
and Signal Processing 2007:21: pp 2560-2574. [18] Khazaee M, Ahmadi H, Omid M, Moosavian A. Fault
[2] Huang B. Detection of abrupt changes of total lest square diagnosis and classification of planetary gearbox of MF285
models and application in fault detection. IEEE Transactions on tractor final drive using Fast Fourier Transform (FFT), Stepwise
Control Systems Technology 2001;9(2): pp 357-367. Backward Selection and support vector machine (SVM)
[3] Mollazade K, Ahmadi H, Omid M, Alimardani R. Vibration- classifier. Elixir Control Engg 2012;43: pp 6974-6977.
Based Fault Diagnosis of Hydraulic Pump of Tractor Steering [19] Ahmadi H, Moosavian A. Fault Diagnosis of Journal-
System by Using Energy Technique. Modern Applied Science Bearing of Generator Using Power Spectral Density and Fault
2009;3(6): pp 59-66. Probability Distribution Function. Innovative Computing
[4] Selection Guide with Bound and Hydro Computer-Assisted Technology, Communications in Computer and Information
Sleeve Bearing Design. Cast Copper Alloy Sleeve Bearings. Science 2011;241(1): pp 30-36.
Copper Development Association. [20] Wowk V. Machinery vibration: measurement and analysis.
[5] Ebrahimi E. Vibration analysis for fault diagnosis of rolling McGraw-Hill Publishers. 1991: pp 64-69.
element bearings. Journal of American Science 2012;8(2): pp [21] Mollazade K, Ahmadi H, Omid M, Alimardani R. An
331-336. intelligent combined method based on power spectral density,
[6] Baydar N, Ball A. A Comparative study of acoustic and decision trees and fuzzy logic for hydraulic pumps fault
vibration signals in detection of gear failures using Wigner-Ville diagnosis. International Journal of Intelligent Systems and
distribution. Mechanical Systems and Signal Processing Technologies 2008;3(4): pp 251-263.
2001;15: pp 1091-1107. [22] Moosavian A, Ahmadi H, Tabatabaeefar A. Condition
[7] Rao MA, Srinivas J, Rama Raju VBV, Kumar KVS. monitoring of engine journal-bearing using power spectral
Coupled torsional-lateral vibration analysis of geared shaft density and support vector machine. Elixir Mech. Engg 2012:43:
systems using mode analysis. Journal of Sound and Vibration pp 6631-6635
2003:261: pp 359-364. [23] Ebrahimi E. Fault diagnosis of Spur gear using vibration
[8] Andrade FA, Esat I, Badi MNM. A new approach to time analysis. Journal of American Science 2012;8(1): pp 133-138.
domain vibration condition monitoring: gear tooth fatigue crack [24] The LOGISTIC Procedure, Chapter 39, Overview, SAS
detection and identification by the Kolmogorov-Smirnov test. OnlineDocä;Version 8: pp 1903-2040.
Journal of Sound and Vibration 2001;240: pp 909-919. [25] Samanta B, Al-Balushi KR, Al-Araimi SA. Artificial neural
[9] Ho D, Randall RB. Optimization of bearing diagnostic networks and support vector machines with genetic algorithm
techniques using simulated and actual bearing fault signals. for bearing fault detection. Engineering Applications of
Mechanical System Signal Process 2000;14: pp763-788. Artificial Intelligence2003;16: pp 657-665.
[10] Hariharan V, Srinivasan PSS. New approach of [26] Y.Song, et.al, IKNN: Informative K-nearest neighbor
classification of rolling element bearing fault using artificial pattern classification. Proc. Springer-Verlag Berlin Heidelberg,
neural network. Journal of Mechanical Engineering 2009;40(2): LNAI 4702, 2007: pp 248–264.
pp119-130.